Coarse-to-Fine Domain Adaptive Semantic Segmentation with Photometric Alignment and Category-Center Regularization

Ma, Haoyu; Lin, Xiangru; Wu, Zifeng; Yu, Yizhou

doi:10.1109/cvpr46437.2021.00404

Cited by 52 publications

(59 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To be specific, GTA5 and SYN-THIA share 19 and 16 common categories with Cityscapes, repectively. On SYNTHIA→Cityscapes, following [32], we consider two different testing protocols: applying all 16 common categories or just a subset consisting of 13 categories for evaluations. Note that we train the model on the Metrics.…”

Section: Methodsmentioning

confidence: 99%

See 1 more Smart Citation

Smoothing Matters: Momentum Transformer for Domain Adaptive Semantic Segmentation

Chen¹,

Yu²,

Guo³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

Section: Methodsmentioning

confidence: 99%

“…Given the past success of Convolutional Neural Networks (CNNs) on many computer vision (CV) tasks [18,29,36], plenty of works [1,6,15,20,21,31,32,43,47,51,54,56] resort to CNNs as the semantic segmentation function f θ . Although these conventional CNN-based backbones obtained decent performance on various benchmarks, recent works (e.g.…”

Section: Introductionmentioning

confidence: 99%

Smoothing Matters: Momentum Transformer for Domain Adaptive Semantic Segmentation

Chen¹,

Yu²,

Guo³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…UDA approaches [3,6,12,17,20,21,23,30,33,41,44] aim at learning domain invariant representations by aligning the distributions of the two domains at feature/output level or at image level. Based on the observation that the source and the target domain share a similar semantic layout, [30,32] rely on adversarial training to align the raw output and entropy distributions respectively.…”

Section: Related Workmentioning

confidence: 99%

“…[6,12,17,38] rely on Cycle-GAN [43] to translate source domain images to the style of the target domain. Two recent works [20,39] bypass the need for training an image translation network by relying on simple Fourier transform and global photometric alignment respectively.…”

Section: Related Workmentioning

confidence: 99%

Beyond Deterministic Translation for Unsupervised Domain Adaptation

Chiou¹,

Panagiotaki²,

Kokkinos³

2022

Preprint

View full text Add to dashboard Cite

Figure 1. Unsupervised Domain Adaptation with stochastic translation: we rely on a content-style separation network to associate a synthetic image from the GTA5 dataset (source) with a distribution of image translations to the target domain. These translations preserve the content signal and adopt the appearance properties of the Cityscapes dataset (target) through randomly sampled style codes, St. We use the resulting images to train a target-domain network tasked with predicting the labels of the respective source-domain image, irrespective of the style variation. Stochasticity in UDA allows the translation networks to generate multiple, sharp outputs that better capture the diversity of the scenes in the target domain, and train the target-domain network with a more representative set of images.

show abstract

“…UDA aims at transferring the knowledge from a labeled source domain to an unlabeled target domain. Existing UDA approaches can be roughly divided into two categories, i.e., aligning domain distributions through adversarial learning [34,36,24,8] and self-training on the target domain [43,39,25,41]. DG focuses on training a robust model with synthetic data, which can generalize well on unseen real-world target data.…”

Section: Related Workmentioning

confidence: 99%

Source-Free Open Compound Domain Adaptation in Semantic Segmentation

Zhao¹,

Zhang²,

Luo³

et al. 2021

Preprint

View full text Add to dashboard Cite

In this work, we introduce a new concept, named source-free open compound domain adaptation (SF-OCDA), and study it in semantic segmentation. SF-OCDA is more challenging than the traditional domain adaptation but it is more practical. It jointly considers (1) the issues of data privacy and data storage and ( 2) the scenario of multiple target domains and unseen open domains. In SF-OCDA, only the source pre-trained model and the target data are available to learn the target model. The model is evaluated on the samples from the target and unseen open domains. To solve this problem, we present an effective framework by separating the training process into two stages: (1) pre-training a generalized source model and (2) adapting a target model with self-supervised learning. In our framework, we propose the Cross-Patch Style Swap (CPSS) to diversify samples with various patch styles in the feature-level, which can benefit the training of both stages. First, CPSS can significantly improve the generalization ability of the source model, providing more accurate pseudo-labels for the latter stage. Second, CPSS can reduce the influence of noisy pseudo-labels and also avoid the model overfitting to the target domain during self-supervised learning, consistently boosting the performance on the target and open domains. Experiments demonstrate that our method produces state-of-the-art results on the C-Driving dataset. Furthermore, our model also achieves the leading performance on CityScapes for domain generalization. * Equal contribution. Preprint. Under review.

show abstract

Coarse-to-Fine Domain Adaptive Semantic Segmentation with Photometric Alignment and Category-Center Regularization

Cited by 52 publications

References 23 publications

Smoothing Matters: Momentum Transformer for Domain Adaptive Semantic Segmentation

Smoothing Matters: Momentum Transformer for Domain Adaptive Semantic Segmentation

Beyond Deterministic Translation for Unsupervised Domain Adaptation

Source-Free Open Compound Domain Adaptation in Semantic Segmentation

Contact Info

Product

Resources

About